MATCHING OF HIGH RESOLUTION SATELLITE IMAGE AND TREE CROWN MAP
Mamoru Kubo and Ken-ichiro Muramoto
School of Electrical and Computer Engineering
Kanazawa University
Kakuma-machi, Kanazawa 920-1192, Japan kubo@ec.t.kanazawa-u.ac.jp, muramoto@t.kanazawa-u.ac.jp
Commission WG IV/9
KEY WORDS: Photogrammetry, Forestry, Matching, Registration, IKONOS
ABSTRACT:
In forest area, there are few landmarks to be ground control points (GCPs) used for registration of satellite images or maps. Additionally,
geographic information from the Global Positioning System (GPS) in field measurement survey is insufficient accuracy to identify
individual tree crowns from satellite image. In this study, we propose the method of identifying individual tree crowns from satellite
image using field measured data. First, in order to obtain the field measured data, we collected several information of individual
trees in the test site. These are the tree stand locations, the distances between the tree trunk and outermost branch in eight directions,
the diameter at breast height, and tree species. Then, using the field measured data, we created the projected on-ground crown map
which has the location and shape of individual trees. The each shape of tree crown is octagonal. Next, we detected the regions of tree
crown from IKONOS panchromatic image. Watershed algorithm was used for image segmentation, based on mathematical morphology
considers gray-scale images to be sets of points in a three-dimensional space, the third dimension being the gray level. The segmented
regions were classified to discriminate tree crown using the feature of spectral signature. Finally, we found out individual tree crowns
related with field measured data from satellite image. Using a GCP by GPS equipment, we performed roughly registration of the
satellite image to the projected on-ground crown map. For each tree crown in the map, we found out the same tree, which has the
highest corresponding possibility to the tree crown in the map, among segmented regions obtained from satellite image. This treeto-tree matching algorithm was performed using the fitness value of the location and octagonal shape of both tree crowns in image
and map. We could obtain the optimum registration by affine transformation of highest fitness value without ground control points.
Consequently, we could identify individual tree crowns from satellite image by image-to-map rectification.
1
INTRODUCTION
Forest composed of many trees has an important role in maintaining environmental conditions suitable for life on the earth.
Satellite remote sensing technology is the effective method for
management and monitoring of forest resources.
In recent years, high spatial resolution satellites were launched,
thereby it is possible to obtain detailed information about earth’s
surface. The IKONOS satellite image can recognize and identify
an individual tree crown, it is suitable to monitor a forest covering
wide-area. In order to obtain forest management inventories at
the stand level, IKONOS satellite images are analyzed instead
of the interpretation of aerial photographs(Gougeon and Leckie,
2006).
To identify tree crown detected from satellite image using field
measured data, we requires high-accuracy image-to-map rectification. However, geographic information from GPS in field
survey is insufficient accuracy to identify individual tree crowns
from satellite image. Additionally, in forest area, there are few
landmarks to be GCPs used for registration of satellite images or
maps.
In this study, we propose the method to identify individual tree
crown from satellite image by image-to-map rectification. This
method is useful for forest management and monitoring.
meter from west to east and 60 meter from north to south. In
this site, there are two flux towers to measure the exchanges of
carbon dioxide between forest and atmosphere. In addition, the
grid of 10 meter mesh is constructed using piles labeled alphabet
and numeric characters. The illustration of this site is shown in
Figure 1.
2.1 Field Measured Data
Field measurement survey was carried out on 28 Octorber 2007.
102 canopy trees with height of 16 to 18 meter were selected in
order to create the projected on-ground map. The relative location
of the tree in this area is acquired by measuring the location in
the labeled block where the tree stands. The following are the
measurement parameters of each canopy tree in this survey:
(1) tree stand location (x, y) in the labeled block;
(2) distances between the tree trunk and outermost branch in
eight directions (N, NE, E, SE, S, SW, W, NW);
(3) diameter at breast height.
The positional information of the flux tower was also recorded.
The illustration of the field survey is shown in Figure 1.
2.2 Satellite Image
2
DATA SET
The Kitasaku test site of this study is located in the deciduous
mixed forest of Nagano prefecture in Japan. This area is 140
1401
The satellite data used in this study is an IKONOS panchromatic
image. The spatial resolution of analysis image is 1 meter by
pixel. It can be recognized and identified an individual tree crown
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008
Figure 3: The illustration of image-to-map rectification.
Figure 1: The illustration of test area. The grid of 10 meter mesh
is constructed using piles labeled alphabet and numeric characters. There are two flux towers to measure the exchanges of carbon dioxide between forest and atmosphere. The field measurement area is 80 meter by 40 meter.
Figure 2: The IKONOS panchromatic image of the study area.
This area is 180 meter by 150 meter. (C)Japan Space Imaging
Co.
Figure 4: Flow diagram of tree-to-tree matching.
this study is octagonal. We obtained the projected on-ground map
from the measurement data of 102 canopy trees.
whose radius is more 2 or 3 meter. The image was acquired on
25 August 2003. Figure 2 shows the IKONOS image of the study
area. The size of image is 180 meter by 150 meter.
3
3.2 Image Segmentation
The IKONOS panchromatic image was segmented using watershed algorithm(Kubo and Muramoto, 2005). Then, the segmented
regions were classified to discriminate tree crowns using the feature of spectral signature.
METHODS
Figure 3 shows the illustration of image-to-map rectification. First,
using the field measured data, we create the projected on-ground
map which has the location and shape of canopy trees. Next, we
detect the region of tree crown from satellite image. Then, we
perform roughly registration of the satellite image and the projected on-ground map. This is initial registration. Finally, performing tree-to-tree matching algorithm(Xiaowei et al., 2006),
we obtain the optimum registration and identify individual tree
crowns.
3.1 The Projected On-Ground Map
The projected on-ground map is a figure that represents the location of canopy tree and the shape. The shape of tree crown in
1402
3.3 Tree-to-Tree Matching
Using the positional information of the east tower by GPS equipment, we performed roughly registration of the satellite image to
the on-ground map. This is initial registration. Then, using affine
transformation, the projected on-ground map is translated, rotated
and scaled in order to find the optimum registration. In each overlap of registration, we performed tree-to-tree matching algorithm
and calculated fitness value. When the fitness value becomes the
maximum, we obtain the optimum parameters of affine transformation for rectifying satellite image to the map coordinate, and
identify tree crowns. Figure 4 shows the flow diagram of tree-totree matching.
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008
Table 1: The affine transformation parameters, range and step for
search
Parameter
Range
Step
Translation:dx, dy
−25.0 ≤ dx, dy ≤ 25.0
0.1 meter
Rotation
:θ
−5.0 ≤ θ ≤ 5.0
0.5 degree
Scaling
:s
−0.90 ≤ s ≤ 1.10
0.01
Figure 6: The projected on-ground map of 102 canopy trees.
Figure 5: The illustration of tree overlap. A[i] is the tree crown in
the projected on-ground map. B[j] is the segmented region from
satellite image.
3.3.1 Initial Registration Using the positional information
of the east tower, we performed roughly registration of the satellite image to map.
3.3.2 Finding The projected on-ground crown map was overlapped to the satellite image using affine transformation. In order to find the optimum registration, we performed tree-to-tree
matching and calculated fitness value in each overlap. The equation of the affine transformation is defined as:
( )
x
y
(
=
cos θ
sin θ
− sin θ
cos θ
)(
s
0
0
s
)( )
x0
y0
(
+
)
dx
dy
(1)
where (x0 , y0 ) is the location of the east tower at initial registration. The affine parameters range of search and step size are
shown in Table 1.
3.3.3 Matching For each tree crown in the map, we find out
the same tree among segmented regions from satellite image,
which has the highest corresponding possibility to the tree crown
in the map. Figure 5 shows illustration of tree overlap.
OL[i][j] =
A[i] ∩ B[j]
A[i] ∩ B[j]
£
A[i]
B[j]
(2)
where A[i]{i = 1 . . . N } is the tree crown in the projected onground map, and B[j]{j = 1 . . . M } is the segmented region
from satellite image. The region B[ki ]{ki = 1 . . . M } of the
highest value is defined as:
OL[i][ki ] ≥ OL[i][j]
for j = 1 . . . M.
(3)
3.3.4 Fitness value The tree-to-tree matching algorithm is performed using the fitness value of the location and octagonal shape
of both tree crowns in the satellite image and the projected onground map.
The fitness value P at each overlap by affine transformation is
defined as:
P (dx, dy, θ, s) =
N
1 ∑
OL[i][ki ].
N
When the fitness value becomes the maximum, we obtain the optimum parameters of affine transformation for rectifying satellite
image to the map coordinate, and identify tree crowns.
4
RESULTS
4.1 The Projected On-Ground Map
The projected on-ground map created from the measurement data
of 102 canopy trees is shown in Figure 6.
4.2 Image Segmentation
The degree of tree overlap is defined as:
√
Figure 7: 911 regions of tree crown were detected by classification after segmentation from satellite image.
(4)
i=1
1403
The tree crowns were detected by classification. Figure 7 shows
the 911 regions of tree crown.
4.3 Tree-to-Tree Matching
The affine transformation parameters at optimum image-to-map
rectification is shown in Table 2. The initial registration of treeto-tree matching and the histogram of tree overlap are shown in
Figure 8. The optimum registration of tree-to-tree matching and
the histogram of tree overlap are shown in Figure 9. The average of the tree overlap at the optimum registration was increased
from 0.437 to 0.509 compared to the initial registration. The locations of the towers in the projected on-ground map overlapped
with the locations of the towers by visual inspection in satellite
image. Using the method in this study, we obtained equivalent
result with the accuracy of image registration by using ground
control points. By the method in this study, the optimum registration were obtained without ground control points. Figure 10
shows rendering the satellite image on the octagonal shapes of
tree crown. Figure 11 shows perspective projection of canopy 3D
model using OpenGL.
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008
Table 2: The affine transformation parameters at optimum imageto-map rectification
Translation:dx, dy
-2.7, 16.2 meter
Rotation
:θ
-1.0 degree
Scaling
:s
1.01
Figure 8: The initial registration of tree-to-tree matching and the
histogram of tree overlap. P = 0.437
Figure 11: Perspective projection of canopy 3D model using
OpenGL.
5
CONCLUSION
In this study, we proposed the method to identify tree crown from
satellite image by image-to-map rectification. The tree-to-tree
matching algorithm was performed using the fitness value of the
location and octagonal shape of both tree crown in satellite image
and field measurement map. We could obtain the optimum registration by affine transformation of highest fitness value without
ground control points.
Furthermore, it became possible to obtain the spectral information such a normalized difference vegetation index (NDVI) from
multi-spectral satellite data, about individual trees. This method
is useful for forest management and monitoring.
ACKNOWLEDGEMENT
Figure 9: The optimum registration of tree-to-tree matching and
the histogram of tree overlap. P = 0.509
This research was supported in part by Grant-in-Aid for Scientific
Research (C) from Japan Society for the Promotion of Science
(No.18580259).
REFERENCES
F.A.Gougeon and D.G.Leckie, 2006. The individual tree crown
approach applied to IKONOS images of a coniferous plantation
area. Photogrammetric Engineering & Remote Sensing, vol.72,
no.11, pp.1287-1297.
Figure 10: Rendering the satellite image on the octagonal shapes
of crown.
Xiaowei Yu, Juha Hyyppä, Antero Kukko, Matti Maltamo, and
Harri Kaatinen, 2006. Change detection techniques for canopy
height growth measurements using airborne laser scanner data.
Photogrammetric Engineering & Remote Sensing, vol.72, no.12,
pp.1339-1348.
M. Kubo. K. Muramoto, 2005. Tree crown detection and classification using forest imagery by IKONOS. IGARSS Proceedings,
vol.6, pp.4358-4361.
1404